Protocol buffers are Google's language-neutral, platform-neutral, extensible mechanism for serializing structured data -- think XML, but smaller, faster, and simpler. You define how you want your data to be structured once. This takes the form of a template that describes the data structure. You use this template to encode your data structure into wire-streams that may be sent-to or read-from your peers. The underlying wire stream is platform independent, lossless, and may be used to interwork with a variety of languages and systems regardless of word size or endianness. Techniques exist to safely extend your data structure without breaking deployed programs that are compiled against the "old" format.
The idea behind Google's Protocol Buffers is that you define your
structured messages using a domain-specific language. This takes the
form of a .proto
source file. You pass this file through a
Google provided tool that generates source code for a target language,
creating an interpreter that can encode/decode your structured data. You
then compile and build this interpreter into your application program.
Depending on the platform, the underlying runtime support is provided by
a Google supplied library that is also bound into your program.
In SWI-Prolog, the wire stream interpreter is embodied in the form of a Definite Clause Grammar (DCG). It has a small underlying C-support library that loads when the Prolog module loads. This implementation does not depend on any code that is provided by Google and thus, is not bound by its license terms.
On the Prolog side, you define your message template as a list of
predefined Prolog terms that correspond to production rules in the DCG.
The process is not unlike specifiying the format of a regular
expression. To encode a message, X
, to wire-stream, Y
,
you pass a grounded template, X
, and a variable, Y
,
to protobuf_message/2. To
decode a wire-stream, Y
, to template, X
, you
pass an ungrounded template, X
, along with a grounded
wire-stream, Y
, to
protobuf_message/2. The
interpreter will unify the unbound variables in the template with values
decoded from the wire-stream.
The wire-stream consists of six primitive payload types, two of which have been deprecated. A primitive in the wire-stream is a multi-byte string that provides three pieces of information: a wire-type, a user-specified tag, and the raw payload. Except for the tag and its wire-type, protobuf payloads are not instantaneously recognizable because the wire-stream contains no payload type information. The interpreter uses the tag to associate the raw payload with a local host type specified by the template. Hence, the message can only be properly decoded using the template that was used to encode it. Note also that the primitive is interpreted according to the needs of a local host. Local word-size and endianness are dealt with at this level.
The following table shows the association between various "host types" used by several peer languages, and the primitives used in the wire-stream:
Prolog Wirestream C++ Java Notes double fixed64 double double integer64 fixed64 int64 long float fixed32 float float integer32 fixed32 int32 int integer varint int32/64 int/long 1, 2 unsigned varint uint32/64 int/long 2, 3 boolean varint bool boolean 2 enum varint int int 2 atom length delimited string String codes length delimited string ByteString string length delimited string String
Notes:
A tag is a small integer that is present in every wire-stream primitive. The tag is the only means that the interpreter has to synchronize the wire-stream with its template. Tags are user defined for each term in each message of the wire-stream. It is important therefore, that they be chosen carefully and in such a way as to not introduce ambiguity.
A protobuf wire-stream is a byte string that is comprised of zero or more of the above multi-byte wire-stream primitives. Templates are lists of Prolog terms. Each term corresponds to a production rule in the DCG. The purpose of the template is to provide a recipe and value set for encoding and decoding a particular message. Each term in the template has an arity of two. The term's functor is the local "host type". Argument 1 is its tag, which must always be ground, and argument 2 is its associated value, which may or may not be ground.
Note: It is an error to attempt to encode a message using a template that is not ground. Decoding a message into a template that has unbound variables has the effect of unifying the variables with their corresponding values in the wire-stream.
Map a Prolog structure to a Protocol Buffer:
command(add(X,Y), Proto) :- Proto = protobuf([atom(1, command), atom(2, add), integer(3, X), integer(4, Y) ]).
Later on:
... prepare X, Y for command ... command(add(X,Y), Proto), protobuf_message(Proto, Msg), ... send the message ...
Proto is the protobuf template. Each template describes exactly one
message. Msg is the wire-stream. If you are interworking with other
systems and languages, then the protobuf templates that you supply to
protobuf_message/2 must
be equivalent to those described in the
.proto
file that is used on the other side.
The protobuf grammar provides a reserved word, optional
,
that indicates that the production rule that it refers to may appear
once or not at all in a protobuf message. Since Prolog has its own means
of alternation, this reserved word is not supported on the Prolog side.
It is anticipated that customary Prolog mechanisms for nondeterminism
(e.g. backtracking) will be used to generate and test alternatives.
It is possible to specify homogeneous vectors of things (e.g. lists
of numbers) using the repeated
attribute. You specify a
repeated field as follows:
repeated(22, float([1,2,3,4])).
The above will cause all four items in the list to be encoded in the wire-stream as IEEE-754 32-bit floating point numbers, all with tag 22. The decoder will aggregate all items in the wire-stream with tag 22 into a list as above.
Notes:
Beware that there is no explicit means to encode an empty set. The
protobuf specification provides that a repeated
field may
match a tag zero or more times. The empty set, while legal, produces no
output on encode. While decoding a repeated
term, failure
to match the specified tag will yield an empty set of the specified host
type.
The protobuf grammar provides a variant of the repeated
field known as "packed." Packed, repeated fields are currently not
supported by our interpreter.
It is possible to embed one protocol buffer specification inside
another using the embedded
term. The following example
shows a vector of numbers being placed in an envelope that contains a
command enumeration.
Enumerations are a compact method of sending tokens from one system
to another. Most occupy only two bytes in the wire-stream. An
enumeration requires that you specify a callable predicate like commands/2,
below. The first argument is an atom specifying the name of token, and
the second is an non-negative integer that specifies the token's value.
These must of course, match a corresponding enumeration in the
.proto
file.
Note: You must expose this predicate to the protobufs module by assigning it explicitly.
protobufs:commands(Key, Value) :- commands(Key, Value). commands(square, 1). commands(decimate, 2). commands(transform, 3). commands(inverse_transform, 4). basic_vector(Type, Proto) :- vector_type(Type, Tag), Proto = protobuf([ repeated(Tag, Type) ]). send_command(Command, Vector, Msg) :- basic_vector(Vector, Proto1), Proto = protobuf([enum(1, commands(Command)), embedded(2, Proto1)]), protobuf_message(Proto, Msg).
Use it as follows:
?- send_command(square, double([1,22,3,4]), Msg). Msg = [8, 1, 18, 36, 17, 0, 0, 0, 0, 0, 0, 240, 63, 17, 0, 0, 0, 0, 0, 0, 54, 64, 17, 0, 0, 0, 0, 0, 0, 8, 64, 17, 0, 0, 0, 0, 0, 0, 16, 64]. ?- send_command(Cmd, V, $Msg). Cmd = square, V = double([1.0, 22.0, 3.0, 4.0]) .
Compatibility Note: The protobuf grammar (protobuf-2.1.0) permits enumerations to assume negative values. This requires them to be encoded as integers. But Google's own Golden Message unit-test framework has enumerations encoded as unsigned. Consequently, parsers that encode them as integers cannot properly parse the Golden Message. So it's probably a good idea to avoid negative values in enumerations. Our parser forbids it anyway.
Using Protocol Buffers, it is quite an easy matter to specify fixed data structures and homogeneous vectors like one might find in languages like C++ and Java. It is however, quite another matter to interwork with these languages when requirements call for working with compound structures, arrays of compound structures, or unstructured collections (e.g. bags) of data.
At bottom, a wire-stream is nothing more than a concatenated stream of primitive wire type strings. As long as you can associate a tag with its host type in advance, you will have no difficulty in decoding the message. You do this by supplying the structure. Tell the parser what is possible and let the parser figure it out on its own, one production at a time.
The following example shows how one can specify a Protocol Buffer message that can deal with variable-length, unstructured bags of numbers:
compound_protobuf(complex(Real, Img), group(12, [double(1, Real), double(2, Img)])). compound_protobuf(float(Val), float(13, Val)). compound_protobuf(double(Val), double(14, Val)). compound_protobuf((Num rdiv Den), group(15, [integer(1, Num), integer(2, Den)])). compound_protobuf(integer(Val), integer(16, Val)). protobuf_bag([], []). protobuf_bag([ Type | More], Msg) :- compound_protobuf(Type, Proto), protobuf_message(protobuf([Proto]), Msg, Msg1), protobuf_bag(More, Msg1), !.
Use it as follows:
?- protobuf_bag([complex(2,3), complex(4,5), complex(6,7), (355 rdiv -113), integer(11)], X). X = [ 99, 9, 0, 0, 0, 0, 0, 0, 0, 64, 17, 0, 0, 0, 0, 0, 0, 8, 64, 100, 99, 9, 0, 0, 0, 0, 0, 0, 16, 64, 17, 0, 0, 0, 0, 0, 0, 20, 64, 100, 99, 9, 0, 0, 0, 0, 0, 0, 24, 64, 17, 0, 0, 0, 0, 0, 0, 28, 64, 100, 123, 8, 198, 5, 16, 225, 1, 124, 128, 1, 22 ]. ?- protobuf_bag(Y, $X). Y = [complex(2.0, 3.0), complex(4.0, 5.0), complex(6.0, 7.0), 355 rdiv -113, integer(11)].
Protocol Buffer Groups provide a means for constructing unitary messages consisting of ad-hoc lists of terms. The following protobuf fragment shows the definition of a group carrying a complex number.
Proto = group(2, [ double(1, Real_part), double(2, Img_part) ]).
Groups have been replaced by embedded messages, which are slightly less expensive to encode.
Performance can be significantly improved using a strategy of precompiling the constant portions of your message. Enumerations for example, are excellent candidates for precompilation. Using protobuf_message/3, the precompiled portion of the message is inserted directly in the wire-stream on encode, and is unified with, and removed from the wire-stream on decode. The following shows how the "send_command" example above, can be converted to precompiled form:
:- dynamic precompiled_message/3. send_precompiled_command(Command, Vector, Msg) :- basic_vector(Vector, Proto1), precompiled_message(commands(Command), Msg, Tail), protobuf_message(protobuf([embedded(3, Proto1)]), Tail). precompile_commands :- abolish(precompiled_message/3), forall(protobufs:commands(Key, _), ( Proto = protobuf([atom(1, command), enum(2, commands(Key))]), protobuf_message(Proto, Msg, Tail), assert(precompiled_message(commands(Key), Msg, Tail)) )), compile_predicates([precompiled_message/3]). * * * :- initialization precompile_commands.
You can extend the parser to support your own compound host types.
These are treated as first class entities by the parser. That is they
can be used either by themselves, or in repeated
and embedded
clauses just as any other host type would be. You do this by hooking
into the parser and adding your own message_sequence
productions. Your hook eventually calls back into the parser with your
substitution/expansion protobuf, which is then embedded in the wire
stream. Recursive structures can be defined this way. A simple example
of a recursive XML like structure is shown below:
:- multifile protobufs:message_sequence/5. protobufs:message_sequence(Type, Tag, Value) --> { my_message_sequence(Type, Value, Proto) }, protobufs:message_sequence(embedded, Tag, Proto), !. % % guard(Type, Value) :- ( nonvar(Value) -> is_of_type(Type, Value); true). my_message_sequence(kv_pair, Key=Value, Proto) :- guard(integer, Value), Proto = protobuf([ atom(31, Key), integer(30, Value)]). my_message_sequence(kv_pair, Key=Value, Proto) :- guard(float, Value), Proto = protobuf([ atom(32, Key), double(30, Value)]). my_message_sequence(kv_pair, Key=Value, Proto) :- guard(atom, Value), Proto = protobuf([ atom(33, Key), atom(30, Value)]). my_message_sequence(xml_element, element(Name, Attributes, Contents), Proto) :- Proto = protobuf([ atom(21, Name), repeated(22, kv_pair(Attributes)), repeated(23, aux_xml_element(Contents))]). my_message_sequence(aux_xml_element, Contents, Proto) :- functor(Contents, element, 3), Proto = protobuf([xml_element(40, Contents)]). my_message_sequence(aux_xml_element, Contents, Proto) :- Proto = protobuf([atom(43, Contents)]). xml_proto([element(space1, [foo='1', bar='2'], [fum, bar, element(space2, [fum=3.1415, bum= -14], ['more stuff for you']), element(space2b, [], [this, is, embedded, also]), to, you])]). test_xml(X, Y) :- Proto = protobuf([repeated(20, xml_element(X))]), protobuf_message(Proto, Y). % And test it: ?- xml_proto(X), test_xml(X,Y), test_xml(Z,Y), Z == X. X = [element(space1, [foo='1', bar='2'], [fum, bar, element(space2, [fum=3.1415, bum= -14], ['more stuff for you'] ), element(space2b, [], [this, is|...] ), to, you])], Y = [162, 1, 193, 1, 170, 1, 6, 115, 112|...], Z = [element(space1, [foo='1', bar='2'], [fum, bar, element(space2, [fum=3.1415, bum= -14], ['more stuff for you'] ), element(space2b, [], [this, is|...] ), to, you])]
In Prolog client:
vector_type(double(_List), 2). vector_type(float(_List), 3). vector_type(integer(_List), 4). vector_type(integer64(_List), 5). vector_type(integer32(_List), 6). vector_type(unsigned(_List), 7). vector_type(codes(_List), 8). vector_type(atom(_List), 9). vector_type(string(_List), 10). vector(Type, B):- vector_type(Type, Tag), Proto = protobuf([ repeated(Tag, Type) ]), protobuf_message(Proto, B).
In the C++ client:
The .proto file: message Vector { repeated double double_values = 2; repeated float float_values = 3; repeated sint32 integer_values = 4; repeated fixed64 integer64_values = 5; repeated fixed32 integer32_values = 6; repeated uint32 unsigned_values = 7; repeated bytes bytes_values = 8; repeated string atom_values = 9; repeated string string_values = 10; }
A typical application might consist of an abstract adapter class along with a collection of concrete subclasses that refine an abstract behavior in order to hide the interaction with the underlying protobuf interpreter. An excerpt of such a class follows:
// Abstract class class PBVector { protected: Vector numberList; private: int serializeToStream(ostream &stream) { return numberList.SerializeToOstream(&stream); } int parseFromStream(istream &stream) { return numberList.ParseFromIstream(&stream); } int serializeToString(string & data) { return numberList.SerializeToString( &data); } int parseFromString(const string &data) { return numberList.ParseFromString(data); } public: friend ostream & operator << (ostream & out, PBVector & a) { int ret = a.serializeToStream(out); if(!ret) throw Exception("error while serializing output stream!"); return out; } friend string & operator << (string & out, PBVector & a) { int ret = a.serializeToString(out); if(!ret) throw Exception("error while serializing output string!"); return out; } friend PBVector & operator << (PBVector & a, istream & cin) { int ret = a.parseFromStream(cin); if(!ret) throw Exception("error while parsing input stream!"); return a; } friend PBVector & operator << (PBVector & a, string & cin) { int ret = a.parseFromString(cin); if(!ret) throw Exception("error while parsing input string!"); return a; } }; // Concrete class class DoublePBVector : public PBVector { public: void add_value(double value) { numberList.add_double_values(value); } double value(int at_index) { return numberList.double_values(at_index); } int size(void) { return numberList.double_values_size(); } };
Here is the C++ client application:
string stuff[] = { "one", "two", "three", "four", "five", }; int main(int argc, char *argv[]) { GOOGLE_PROTOBUF_VERIFY_VERSION; DoublePBVector buffer, buffer1; register int i; try { buffer << cin; // parse a wirestream from a istream & } catch (Exception e) { cerr << "exception: " << e.reason() << "\n"; exit(1); } string serialized_data; serialized_data << buffer; // serialize a buffer to a string buffer1 << serialized_data; // parse the string to another buffer int size = buffer1.size(); // inspect the second instance for(i = 0; i < size; i++) cout << buffer1.value(i) << "\n"; return 0; }
On the Prolog side:
:- meta_predicate ~>(0,0). :- op(950, xfy, ~>). ~>(P, Q) :- setup_call_cleanup(P, (true; fail), assertion(Q)). write_as_proto(Vector) :- vector(Vector, Wirestream), open('tmp99.tmp', write, S, [type(binary)]) ~> close(S), format(S, '~s', [Wirestream]), !. testv1(V) :- read_file_to_codes('tmp99.tmp', Codes, [type(binary)]), vector(V, Codes).
Run the Prolog side:
?- X is pi, write_as_proto(double([-2.2212, -7.6675, X, 0, 1.77e-9, 2.54e222])). X = 3.14159. ?- testv1(Vector). Vector = double([-2.2212, -7.6675, 3.14159, 0.0, 1.77e-09, 2.54e+222]) ?-
Make and run the C++ side:
jeffr@jeffr-laptop:~/test/protob$ make foo protoc --cpp_out=. pb-vector.proto g++ -c -Wall -O3 pb-vector.pb.cc g++ -c -Wall -O3 foo.cpp g++ -o foo pb-vector.pb.o foo.o -lprotobuf jeffr@jeffr-laptop:~/test/protob$ foo <tmp99.tmp -2.2212 -7.6675 3.14159 0 1.77e-09 2.54e+222
Protocol buffers are Google's language-neutral, platform-neutral, extensible mechanism for serializing structured data -- think XML, but smaller, faster, and simpler. You define how you want your data to be structured once. This takes the form of a template that describes the data structure. You use this template to encode and decode your data structure into wire-streams that may be sent-to or read-from your peers. The underlying wire stream is platform independent, lossless, and may be used to interwork with a variety of languages and systems regardless of word size or endianness. Techniques exist to safely extend your data structure without breaking deployed programs that are compiled against the "old" format.
The idea behind Google's Protocol Buffers is that you define your
structured messages using a domain-specific language and tool set. In
SWI-Prolog, you define your message template as a list of predefined
Prolog terms that correspond to production rules in the Definite Clause
Grammar (DCG) that realizes the interpreter. Each production rule has an
equivalent rule in the protobuf grammar. The process is not unlike
specifiying the format of a regular expression. To encode a template to
a wire-stream, you pass a grounded template, X
, and
variable, Y
, to
protobuf_message/2. To
decode a wire-stream, Y
, you pass an ungrounded template, X
,
along with a grounded wire-stream, Y
, to
protobuf_message/2. The
interpreter will unify the unbound variables in the template with values
decoded from the wire-stream.
For an overview and tutorial with examples, see protobufs_overview.txt
.
Examples of usage may also be found by inspecting test_protobufs.pl
.
Template | is a protobuf grammar specification. On decode, unbound variables in the Template are unified with their respective values in the Wire_stream. On encode, Template must be ground. |
Wire_stream | is a code list that was generated by a protobuf encoder using a equivalent template. |